Visual vocabularies for category-level object recognition

  1. López Sastre, Roberto Javier
Supervised by:
  1. Saturnino Maldonado Bascón Director

Defence university: Universidad de Alcalá

Fecha de defensa: 18 May 2010

Committee:
  1. Francisco López Ferreras Chair
  2. Manuel Rosa Zurera Secretary
  3. Nicolás Ruiz Reyes Committee member
  4. Antonio Fernández Caballero Committee member
  5. Alessandro Foi Committee member
Department:
  1. Teoría de la Señal y Comunicaciones

Type: Thesis

Abstract

This thesis focuses on the study of visual vocabularies for category-level object recognition. Specifically, we state novel approaches for building visual codebooks. Our aim is not just to obtain more discriminative and more compact visual codebooks, but to bridge the gap between visual features and semantic concepts. A novel approach for obtaining class representative visual words is presented. It is based on a maximisation procedure, i. e. the Cluster Precision Maximisation (CPM), of a novel cluster precision criterion, and on an adaptive threshold refinement scheme for agglomerative clustering algorithms based on correlation clustering techniques. The objective is to increase the vocabulary compactness while at the same time improve the recognition rate and further increase the representativeness of the visual words. Moreover, we describe a novel clustering aggregation based approach for building efficient and semantic visual vocabularies. It consist of a novel framework for incorporating neighboring appearances of local descriptors into the vocabulary construction, and a rigorous approach for adding meaningful spatial coherency among the local features into the visual codebooks. We also propose an efficient high-dimensional data clustering algorithm, the Fast Reciprocal Nearest Neighbours (Fast-RNN). Our approach, which is a speeded up version of the standard RNN algorithm, is based on the projection search paradigm. Finally, we release a new database of images called Image Collection of Annotated Real-world Objects (ICARO), which is especially designed for evaluating category-level object recognition systems. An exhaustive comparison of ICARO with other well-known datasets used within the same context is carried out. We also propose a benchmark for both object classification and detection.